AI fans have been buzzing since April when whispers about Anthropic’s Claude Mythos Preview first surfaced – a model the company claimed could detect enough security flaws to potentially cripple the internet if left unchecked. That hype led Anthropic to deliberately delay the launch.
On Tuesday, Anth{…} released the inaugural Mythos‑family model that anyone can access: Claude Fable 5.
In a detailed blog post, Anthropic explained why Fable 5 represents its most sophisticated AI yet still respects safety constraints. The model is said to deliver better assistance for software engineering and knowledge‑work tasks, and it shows improved comprehension of images and other non‑text inputs. Notably, it managed to beat Pokémon FireRed using a vision‑only AI setup, something earlier versions could not achieve.
While many are eager to explore the capabilities of the Mythos series, these models demand substantial compute resources, so Anthropic isn’t offering them at a low price. In a few weeks, the company plans to start charging subscribers an extra fee for using Fable 5.
Is Claude Fable 5 safe?
Anthropic positions Claude Fable 5 as a safer counterpart to Claude Mythos 5, which is limited to trusted cybersecurity and software experts via Anthropic’s Project Glasswing program. The Fable variant incorporates additional cybersecurity safeguards that Mythos lacks.
«Fable 5 rejected every harmful single‑turn request related to planning a cyber‑attack, exploit development, or defense evasion,» the company reported, adding that this held true even when the prompts employed any of 30 known public jailbreak techniques.
AI‑driven bioterrorism—a chilling phrase and an even more unsettling prospect—has been a growing worry for leading AI labs. Consequently, Anthropic has disabled Fable 5 from answering most biology and chemistry queries for now, routing those to another model, Opus 4.8.
The firm also disclosed that it is consulting with the U.S. government about the rollout of Claude Mythos 5. Mythos played a key role in the White House’s recent push for AI firms to submit new models for governmental review before release, though that requirement is not yet in force.
In a policy shift, business users of Claude must now accept a 30‑day data‑retention rule. Anthropic says retained data will be used to defend against future cyber threats and AI misuse, not to train new models.
How to access Claude Fable 5
Fable 5 is currently available to Claude subscribers, but Anthropic says the rollout will be «more conservatively, in stages.» The model will be offered to paid plans—Pro, Max, Team, and Enterprise—until June 22. Starting June 23, Fable 5 will disappear from the subscription options list.
After the initial window, users can still run Fable 5 by spending usage credits, a «pay‑as‑you‑go» system that lets you continue using Claude beyond your standard quota. Many enterprise plans already include such credits, but if you run out or have none, extra charges will appear on your next invoice.
Developers can call Fable 5 via the Claude API at $10 per million input tokens and $50 per million output tokens.
The unusual rollout likely stems from capacity concerns: selecting Fable 5 in Claude shows a pop‑up noting it consumes twice the compute of Anthropic’s Opus models. Anthropic expects strong demand, so the brief availability window helps keep the service stable.
Anthropic says it eventually hopes to integrate Fable 5 into its regular paid plans, though no firm timeline has been announced.
For additional reading, see Apple Intelligence’s WWDC redesign and the roundup of top AI chatbots.
